Strong and Precise Modulation of Human Percepts
via Robustified ANNs (NeurIPS 2023)

Official implementation & data for:

Strong and Precise Modulation of Human Percepts via Robustified ANNs
Guy Gaziv*, Michael J. Lee*, James J. DiCarlo

TL;DR
Robustified ANNs reliably discover image perturbations in the "human-presumed-stable" pixel-budget regime that can strongly disrupt human behavior, or can precisely drive human perceptual state transitions to prescribed target states.

Table of Contents

• Setup
  • Data & Results
  • Environment
  • Test code
• Paper results
• Generating images
• Training models
• Behavioral experiments
• Acknowledgments
• Citation

Setup

Environment

Clone this repo and create a designated conda environment using the env.yml file:

cd Wormholes
conda env create -n wormholes -f env.yml
conda activate wormholes

Recommended: for quick environment solving use libmamba-solver or mamba directly:

conda install -y -c conda-forge mamba

(Then just use mamba instead of conda in the above calls)

Data & Results

We make all our data and results publicly available. This includes source data, model checkpoints, perturbed images with their metadata, and the MTurk behavioral reports (anonymized). For convenience, we include download scripts for each component under ./scripts/download. The following convenience script sets up all data automatically.

./scripts/download/all.sh

Finally, download ILSVRC and place a link to it under data/:

ln -s <ILSVRC-FOLDER> data/

Notably, for completeness, some of the checkpoints are direct copies of those published in robustness. 'v2' versions of models are equivalent to the non-'v2' versions but initialized and trained using a different seed.

Test code runs successfully

export CUDA_VISIBLE_DEVICES=<CUDA_ID>
pytest -sv tests/test_gen_smoketest.py

Viewing paper results

We include two notebooks for visualization for results under notebooks/:

• images_view.ipynb -- visualizing image perturbation results with their metadata. Note that this notebook assumes having ILSVRC validation set, as .data/ilsvrc/val.
• results_figures.ipynb -- replicating all paper figures.

To work with these notebooks, you might need to install the jupyter kernel of the environment:

python -m ipykernel install --user --name "wormholes" --display-name wormholes

Perturbing images

All our paper related image perutbation experiments are implemented under wormholes/perturb/. To run an experiment gen_v<X>.py:

export CUDA_VISIBLE_DEVICES=<CUDA_ID>
./scripts/gen_data.sh <X>

Results will output under results/cache/ (note they might already exist there if you downloaded them). Lastly you will need to merge the generated metadata to a single file:

python scripts/combine_results.py --gen_version <X>

Training models

Scripts for training models and surrogate models, similar to those in the paper (albeit from a different seed), are included under scripts/train/.

Behavioral experiments (psychophysics)

Follow instructions under psych/.

Acknowledgments

This code borrows from robustness.

Citation

If you find this repository useful, please consider giving a star ⭐️ and citation:

@article{Gaziv2023,
  title  = {Robustified ANNs Reveal Wormholes Between Human Category Percepts},
  author = {Guy Gaziv and Michael J. Lee and James J. DiCarlo},
  url    = {https://arxiv.org/abs/2308.06887},
  year   = {2023},
}